Load libraries and data

easypackages::libraries("here","ggplot2","caret","e1071","pheatmap","reshape2","NbClust","grid","patchwork","readxl","patchwork","WGCNA","psych","nlme")
source(here("code","ndar_functions.R"))
source(here("code","euaims_functions.R"))
source(here("code","get_ggColorHue.R"))

options(stringsAsFactors = FALSE)

# Z-score threshold to use for subtyping
z_thresh = 0.7

codepath = here("code")
datapath = here("data")
figpath = here("figures")
resultpath = here("results","ndar")
plotpath = here("plots","ndar")

# read in data
Dverbal_Discovery = read.csv(file.path(datapath,"tidy_verbal_disc.csv"))
Dverbal_Replication = read.csv(file.path(datapath,"tidy_verbal_rep.csv"))

Subtyping using Z-score of the difference between SC and RRB

make_subtype <- function(data2use, z_thresh, mean2use=NULL, sd2use=NULL){
  # compute difference score
  vars2use = c("dbaes_atotal","dbaes_btotal")
  diff_score = data2use[,vars2use[1]] - data2use[,vars2use[2]]
  
  # compute mean and sd if necessary
  if (is.null(mean2use)){
    mean2use = mean(diff_score)
  } # if (is.null(mean2use))
  
  if (is.null(sd2use)){
    sd2use = sd(diff_score)
  } # if (is.null(sd2use))
  
  # compute z-score
  data2use$z_ds = (diff_score - mean2use)/sd2use
  
  # make subtype factor
  data2use$z_ds_group = "SC_equal_RRB"
  data2use$z_ds_group[data2use$z_ds>z_thresh] = "SC_over_RRB"
  data2use$z_ds_group[data2use$z_ds<(z_thresh*-1)] = "RRB_over_SC"
  data2use$z_ds_group = factor(data2use$z_ds_group)
  return(data2use)
  
} # function make_subtype

vars2use = c("dbaes_atotal","dbaes_btotal")

# compute Discovery mean and sd
ds_disc = Dverbal_Discovery[,vars2use[1]] - Dverbal_Discovery[,vars2use[2]]
mean2use = mean(ds_disc) 
sd2use = sd(ds_disc)

Dverbal_Discovery = make_subtype(data2use = Dverbal_Discovery,
                                 z_thresh = z_thresh,
                                 mean2use = mean2use,
                                 sd2use = sd2use)

# compute Replication mean and sd
ds_rep = Dverbal_Replication[,vars2use[1]] - Dverbal_Replication[,vars2use[2]]
# mean2use = mean(ds_rep) 
# sd2use = sd(ds_rep)

Dverbal_Replication = make_subtype(data2use = Dverbal_Replication,
                                 z_thresh = z_thresh,
                                 mean2use = mean2use,
                                 sd2use = sd2use)

Make scatterplots with difference score Z subtypes in different colors

maxScores = c(3,4)

p_disc = ggplot(data = Dverbal_Discovery, aes(x = dbaes_atotal, y = dbaes_btotal, colour = z_ds_group)) + geom_point() + xlab("SC") + ylab("RRB") + ylim(0,1) + xlim(0,1) + ggtitle("NDAR Discovery")
p1_top_left = p_disc + guides(colour=FALSE)
ggsave(filename = file.path(plotpath, sprintf("final_NDAR_Disc_scatterplot_z%s.pdf",as.character(z_thresh))), plot = p1_top_left)
p_disc

table(Dverbal_Discovery$z_ds_group)
## 
##  RRB_over_SC SC_equal_RRB  SC_over_RRB 
##          211          477          201
cor_res = cor.test(Dverbal_Discovery$dbaes_atotal, Dverbal_Discovery$dbaes_btotal)
cor_res
## 
##  Pearson's product-moment correlation
## 
## data:  Dverbal_Discovery$dbaes_atotal and Dverbal_Discovery$dbaes_btotal
## t = 6.6829, df = 887, p-value = 4.134e-11
## alternative hypothesis: true correlation is not equal to 0
## 95 percent confidence interval:
##  0.1554332 0.2806578
## sample estimates:
##       cor 
## 0.2189469
p_rep = ggplot(data = Dverbal_Replication, aes(x = dbaes_atotal, y = dbaes_btotal, colour = z_ds_group)) + geom_point() + xlab("SC") + ylab("RRB") + ylim(0,1) + xlim(0,1) +  ggtitle("NDAR Replication")
p2_bottom_left = p_rep + guides(colour=FALSE)
ggsave(filename = file.path(plotpath, sprintf("final_NDAR_Rep_scatterplot_z%s.pdf",as.character(z_thresh))), plot = p2_bottom_left)
p_rep

table(Dverbal_Replication$z_ds_group)
## 
##  RRB_over_SC SC_equal_RRB  SC_over_RRB 
##          220          485          185
cor_res = cor.test(Dverbal_Replication$dbaes_atotal, Dverbal_Replication$dbaes_btotal)
cor_res
## 
##  Pearson's product-moment correlation
## 
## data:  Dverbal_Replication$dbaes_atotal and Dverbal_Replication$dbaes_btotal
## t = 7.1459, df = 888, p-value = 1.864e-12
## alternative hypothesis: true correlation is not equal to 0
## 95 percent confidence interval:
##  0.1700824 0.2943934
## sample estimates:
##       cor 
## 0.2331903

Run supervised model with Discovery as training and Replication as Test

# run validation
# make subtypes using z-scores computed from the mean and sd of the training set
train_data = Dverbal_Discovery
test_data = Dverbal_Replication

mean2use = mean(train_data[,vars2use[1]] - train_data[,vars2use[2]])
sd2use = sd(train_data[,vars2use[1]] - train_data[,vars2use[2]])
tmp_train = make_subtype(data2use = train_data,
                         z_thresh = z_thresh,
                         mean2use = mean2use,
                         sd2use = sd2use)

# mean2use = mean(test_data[,vars2use[1]] - test_data[,vars2use[2]])
# sd2use = sd(test_data[,vars2use[1]] - test_data[,vars2use[2]])
tmp_test = make_subtype(data2use = test_data,
                        z_thresh = z_thresh,
                        mean2use = mean2use,
                        sd2use = sd2use)
# compute model
mod2use = svm(x = tmp_train[,vars2use], y = tmp_train$z_ds_group)
pred_labels = predict(mod2use, tmp_test[,vars2use])
confmat = table(tmp_test$z_ds_group,pred_labels)
acc = (confmat[1,1]+confmat[2,2]+confmat[3,3])/length(pred_labels)

# plot confusion matrix
setHook("grid.newpage", function() pushViewport(viewport(x=1,y=1,width=0.9, height=0.9, name="vp", just=c("right","top"))), action="prepend")
pheatmap(confmat/rowSums(confmat)*100, display_numbers = TRUE, color = colorRampPalette(c('white','red'))(100), cluster_rows = FALSE, cluster_cols = FALSE, fontsize_number = 15,labels_row = c("RRB>SC","SC=RRB","SC>RRB"),labels_col = c("RRB>SC","SC=RRB","SC>RRB"),angle_col=90)
setHook("grid.newpage", NULL, "replace")
grid::grid.text("Actual Labels", y=-0.07, gp=gpar(fontsize=16))
grid::grid.text("Predicted Labels", x=-0.07, rot=90, gp=gpar(fontsize=16))

# show accuracy
true_accuracy = acc
true_accuracy
## [1] 0.9719101

Permute subtype labels to examine how well supervised model performs

nperm = 10000
# make subtypes using z-scores computed from the mean and sd of the training set
train_data = Dverbal_Discovery
test_data = Dverbal_Replication
mean2use = mean(train_data[,vars2use[1]] - train_data[,vars2use[2]])
sd2use = sd(train_data[,vars2use[1]] - train_data[,vars2use[2]])
tmp_train = make_subtype(data2use = train_data,
                         z_thresh = z_thresh,
                         mean2use = mean2use,
                         sd2use = sd2use)

# mean2use = mean(test_data[,vars2use[1]] - test_data[,vars2use[2]])
# sd2use = sd(test_data[,vars2use[1]] - test_data[,vars2use[2]])
tmp_test = make_subtype(data2use = test_data,
                        z_thresh = z_thresh,
                        mean2use = mean2use,
                        sd2use = sd2use)

acc = vector(length = nperm)
for (iperm in 1:nperm){
  # set seed for reproducibility
  set.seed(iperm)
  # compute model
  permuted_labels = sample(tmp_train$z_ds_group)
  mod2use = svm(x = tmp_train[,vars2use], y = permuted_labels)
  pred_labels = predict(mod2use, tmp_test[,vars2use])
  confmat = table(tmp_test$z_ds_group,pred_labels)
  acc[iperm] = (confmat[1,1]+confmat[2,2]+confmat[3,3])/length(pred_labels)
} # for (iperm in 1:nperm)

df2plot = data.frame(Accuracy = acc)
p = ggplot(data = df2plot, aes(x = Accuracy)) + geom_histogram() + geom_vline(xintercept=true_accuracy)
p

# compute p-value
pval = sum(c(true_accuracy,acc)>=true_accuracy)/(nperm+1)
pval
## [1] 9.999e-05

Plot difference score Z subtypes in Discovery set

maxScores = c(3,4)

# Discovery - make plot with all individuals shown as lines
df2use = Dverbal_Discovery[,c("subjectkey",adi_total_vars2use)]
df2use$subgrp = factor(tmp_train$z_ds_group)
df2use = data.frame(df2use)
df2use$subjectkey = factor(df2use$subjectkey)
df2use$SC = df2use$dbaes_atotal
df2use$RRB = df2use$dbaes_btotal

df4plot = melt(df2use,
               id.vars = c("subjectkey","subgrp"), 
               measure.vars = c("SC","RRB"))

p = ggplot(data = df4plot, aes(x = variable, 
                               y = value, 
                               colour = subgrp, 
                               group = subjectkey)) + facet_grid(. ~ subgrp)
p = p + geom_point(shape=1) + geom_line(alpha = 0.2) + ylim(0,1) + guides(color=FALSE)
p = p + ylab("Percent Severity") + xlab("ADI-R subscale")
p3_middle_top = p + guides(colour=FALSE)
ggsave(filename = file.path(plotpath, sprintf("final_NDAR_Disc_jitterplot_z%s.pdf",as.character(z_thresh))), plot = p3_middle_top)
p

Plot difference score Z subtypes in Replication set

maxScores = c(3,4)

# Replication - make plot with all individuals shown as lines
df2use = Dverbal_Replication[,c("subjectkey",adi_total_vars2use)]
df2use$subgrp = factor(tmp_test$z_ds_group)
df2use = data.frame(df2use)
df2use$subjectkey = factor(df2use$subjectkey)
df2use$SC = df2use$dbaes_atotal
df2use$RRB = df2use$dbaes_btotal

df4plot = melt(df2use,
               id.vars = c("subjectkey","subgrp"), 
               measure.vars = c("SC","RRB"))

p = ggplot(data = df4plot, aes(x = variable, 
                               y = value, 
                               colour = subgrp, 
                               group = subjectkey)) + facet_grid(. ~ subgrp)
p = p + geom_point(shape=1) + geom_line(alpha = 0.2) + ylim(0,1) + guides(color=FALSE)
p = p + ylab("Percent Severity") + xlab("ADI-R subscale")
p4_middle_bottom = p + guides(colour=FALSE)
ggsave(filename = file.path(plotpath, sprintf("final_NDAR_Rep_jitterplot_z%s.pdf",as.character(z_thresh))), plot = p4_middle_bottom)
p

Apply NDAR subtypes to EU-AIMS LEAP data

# make SC and RRB percentages since EU-AIMS data is specified as percentages
Dverbal = read.csv(file.path(datapath,"tidy_verbal.csv"))
euaims_data = read.csv(file.path(datapath,"tidy_euaims.csv"))
mask1 = euaims_data$Diagnosis=="ASD" 
mask2 =  (is.na(euaims_data$A1_pct_severity) | is.na(euaims_data$A2_pct_severity) | is.na(euaims_data$A3_pct_severity) | is.na(euaims_data$B1_pct_severity) | is.na(euaims_data$B2_pct_severity) | is.na(euaims_data$B3_pct_severity) | is.na(euaims_data$B4_pct_severity)) 
euaims_data = subset(euaims_data, (mask1 & !mask2))


Dverbal[,vars2use[1]] = Dverbal[,vars2use[1]]
Dverbal[,vars2use[2]] = Dverbal[,vars2use[2]]
euaims_data[,vars2use[1]] = (euaims_data$A1_pct_severity + 
                               euaims_data$A2_pct_severity + 
                               euaims_data$A3_pct_severity)/3
euaims_data[,vars2use[2]] = (euaims_data$B1_pct_severity + 
                               euaims_data$B2_pct_severity + 
                               euaims_data$B3_pct_severity + 
                               euaims_data$B4_pct_severity)/4

train_data = Dverbal
test_data = euaims_data

mean2use = mean(train_data[,vars2use[1]] - train_data[,vars2use[2]], na.rm=TRUE)
sd2use = sd(train_data[,vars2use[1]] - train_data[,vars2use[2]], na.rm=TRUE)
c(mean2use, sd2use)
## [1] -0.01045243  0.19482749
tmp_train = make_subtype(data2use = train_data,
                         z_thresh = z_thresh,
                         mean2use = mean2use,
                         sd2use = sd2use)

tmp_test = make_subtype(data2use = test_data,
                        z_thresh = z_thresh,
                        mean2use = mean2use,
                        sd2use = sd2use)

# compute model
mod2use = svm(x = tmp_train[,vars2use], y = tmp_train$z_ds_group)
pred_labels = predict(mod2use, tmp_test[,vars2use])
confmat = table(tmp_test$z_ds_group,pred_labels)
acc = (confmat[1,1]+confmat[2,2]+confmat[3,3])/length(pred_labels)

tmp_test$svm_pred_labels = pred_labels

# plot confusion matrix
setHook("grid.newpage", function() pushViewport(viewport(x=1,y=1,width=0.9, height=0.9, name="vp", just=c("right","top"))), action="prepend")
pheatmap(confmat/rowSums(confmat)*100, display_numbers = TRUE, color = colorRampPalette(c('white','red'))(100), cluster_rows = FALSE, cluster_cols = FALSE, fontsize_number = 15,labels_row = c("RRB>SC","SC=RRB","SC>RRB"),labels_col = c("RRB>SC","SC=RRB","SC>RRB"),angle_col=90)
setHook("grid.newpage", NULL, "replace")
grid::grid.text("Actual Labels", y=-0.07, gp=gpar(fontsize=16))
grid::grid.text("Predicted Labels", x=-0.07, rot=90, gp=gpar(fontsize=16))

# scatterplot
p1 = ggplot(data = tmp_train, aes(x = dbaes_atotal, y = dbaes_btotal, colour = factor(z_ds_group))) + geom_point() + xlab("Social-Communication") + ylab("Restricted Repetitive Behaviors") + ylim(0,0.8) + xlim(0,0.8) + ggtitle("NDAR ALL")
p1

p2 = ggplot(data = tmp_test, aes(x = dbaes_atotal, y = dbaes_btotal, colour = factor(z_ds_group))) + geom_point() + xlab("Social-Communication") + ylab("Restricted Repetitive Behaviors") + ylim(0,0.8) + xlim(0,0.8) + ggtitle("EU-AIMS with Groups from NDAR ALL")
p2

p3 = ggplot(data = tmp_test, aes(x = dbaes_atotal, y = dbaes_btotal, colour = factor(pred_labels))) + geom_point() + xlab("SC") + ylab("RRB") + ylim(0,0.8) + xlim(0,0.8) + ggtitle("EU-AIMS")
p5_bottom_right = p3 + guides(colour=FALSE)
ggsave(filename = file.path(plotpath, sprintf("final_EUAIMS_scatterplot_z%s.pdf",as.character(z_thresh))), plot = p5_bottom_right)
p3

# # write out EU-AIMS LEAP data with subgroups defined by NDAR All
write.csv(tmp_test, file.path(datapath, sprintf("tidy_euaims_NDAR_subtypes_diffscore_z%s.csv",as.character(z_thresh))))

Make final plot

p_final = p1_top_left + p3_middle_top + plot_spacer() + p2_bottom_left + p4_middle_bottom + p5_bottom_right + plot_layout(nrow=3, ncol=3, widths = c(4,4,4), heights = c(8,8,8))
ggsave(filename = file.path(plotpath, sprintf("final_NDAR_EUAIMS_subtypes_plot_z%s.pdf",as.character(z_thresh))), plot = p_final)
p_final

Integrate subtypes with rest of EU-AIMS LEAP data

euaims_data = read.csv(file.path(datapath,"tidy_euaims.csv"))
td_df = subset(euaims_data, euaims_data$Diagnosis=="TD",select=2:6)
td_df$A_pct_severity = as.numeric(NA)
td_df$B_pct_severity = as.numeric(NA)
td_df$subgrp = "TD"

cols2use = c("subid","age","Centre","Schedule","Diagnosis","dbaes_atotal","dbaes_btotal","svm_pred_labels")
tmp_asd = tmp_test[,cols2use]
colnames(tmp_asd)[6] = "A_pct_severity"
colnames(tmp_asd)[7] = "B_pct_severity"
colnames(tmp_asd)[8] = "subgrp"
asd_df = tmp_asd

all_data = rbind(td_df,asd_df)

fname = "/Users/mlombardo/Dropbox/euaims/data/rsfmri_preproc/euaims_preproc.xlsx"
pp_data = read_excel(fname)
mask = pp_data$notes=="ok"
pp_data = subset(pp_data, mask)

asd_df = merge(pp_data[,c("subid","sex")],asd_df, by = "subid")
td_df = merge(pp_data[,c("subid","sex")],td_df, by = "subid")

all_data = rbind(td_df,asd_df)

data2write = merge(pp_data, all_data, by = "subid")
data2write$age = data2write$age.x
data2write$sex = data2write$sex.y
print(table(data2write$Schedule, data2write$subgrp))
##    
##     RRB_over_SC SC_equal_RRB SC_over_RRB TD
##   A           7           37          36 78
##   B           1           45          39 83
##   C           4           29          30 59
##   D           1           10          27 23
print(table(data2write$Centre, data2write$subgrp))
##                
##                 RRB_over_SC SC_equal_RRB SC_over_RRB TD
##   CAMBRIDGE               4           27          11 29
##   KINGS_COLLEGE           7           49          50 78
##   MANNHEIM                0            0           0 34
##   NIJMEGEN                2           36          55 64
##   UTRECHT                 0            9          16 38
print(table(data2write$sex, data2write$subgrp))
##         
##          RRB_over_SC SC_equal_RRB SC_over_RRB  TD
##   Female           5           31          35  88
##   Male             8           90          97 155
#DSM-5 - find best split that balances participants across sites
a = findBestSplit(asd_df, seed_range = c(172342)) #,300001:500000))
## [1] 172342
best_seeds = a$seed[!is.na(a$discrepancy) & a$discrepancy==min(a$discrepancy, na.rm = TRUE)]
print(best_seeds)
## [1] 172342
# Split datasets -------------------------------------------------------------
rngSeed = best_seeds[1]

# split Schedule A dataset
dset2use = subset(asd_df, asd_df$Schedule=="A")
tmp_d = SplitDatasetsBySex(dset2use, rngSeed = rngSeed)
A_Discovery = tmp_d[[2]]
A_Replication = tmp_d[[1]]

# split Schedule B dataset
dset2use = subset(asd_df, asd_df$Schedule=="B")
tmp_d = SplitDatasetsBySex(dset2use, rngSeed = rngSeed)
B_Discovery = tmp_d[[2]]
B_Replication = tmp_d[[1]]

# split Schedule C dataset
dset2use = subset(asd_df, asd_df$Schedule=="C")
tmp_d = SplitDatasetsBySex(dset2use, rngSeed = rngSeed)
C_Discovery = tmp_d[[2]]
C_Replication = tmp_d[[1]]

# split Schedule D dataset
dset2use = subset(asd_df, asd_df$Schedule=="D")
tmp_d = SplitDatasetsBySex(dset2use, rngSeed = rngSeed)
D_Discovery = tmp_d[[2]]
D_Replication = tmp_d[[1]]

df_Disc = rbind(A_Discovery, B_Discovery, C_Discovery, D_Discovery)
df_Rep = rbind(A_Replication, B_Replication, C_Replication, D_Replication)

a = table(df_Disc$Schedule, df_Disc$Centre); print(a)
##    
##     CAMBRIDGE KINGS_COLLEGE NIJMEGEN UTRECHT
##   A         6            17       10       7
##   B         9            16       14       3
##   C         6             9       14       2
##   D         0            10       10       0
b = table(df_Rep$Schedule, df_Rep$Centre); print(b)
##    
##     CAMBRIDGE KINGS_COLLEGE NIJMEGEN UTRECHT
##   A         7            18       10       5
##   B         8            17       13       5
##   C         6            10       13       3
##   D         0             9        9       0
print(a-b)
##    
##     CAMBRIDGE KINGS_COLLEGE NIJMEGEN UTRECHT
##   A        -1            -1        0       2
##   B         1            -1        1      -2
##   C         0            -1        1      -1
##   D         0             1        1       0
print(sum(rowSums(abs(a-b))))
## [1] 14
data2write$dataset = NA
mask  = is.element(data2write$subid,df_Disc$subid)
data2write[mask,"dataset"] = "Discovery"
mask  = is.element(data2write$subid,df_Rep$subid)
data2write[mask,"dataset"] = "Replication"

asd_Disc = subset(data2write, data2write$dataset=="Discovery" & data2write$Diagnosis=="ASD")
asd_Rep = subset(data2write, data2write$dataset=="Replication" & data2write$Diagnosis=="ASD")

# find which seed gives best TD age-match -------------------------------------
seeds = 1:1000
pvals = data.frame(matrix(nrow = length(seeds), ncol = 2))
for (i in 1:length(seeds)) {
  res = findTDAgeMatch(data2write, seed_range = c(seeds[i],seeds[i]))
  td_Disc_matched = res[[2]]
  td_Rep_matched = res[[1]]
  tres = t.test(td_Disc_matched$age, asd_Disc$age)
  pvals[i,1] = tres$p.value
  tres = t.test(td_Rep_matched$age, asd_Rep$age)
  pvals[i,2] = tres$p.value
  #print(i)
}
a = sort.int(pvals[,1], decreasing = TRUE, index.return = TRUE)
b=pvals[a$ix,]

seed2use = 929
res = findTDAgeMatch(data2write, seed_range = c(seed2use,seed2use))
td_Disc_matched = res[[2]]
td_Rep_matched = res[[1]]

mask = is.element(data2write$subid, td_Disc_matched$subid)
data2write$dataset[mask] = "Discovery"

mask = is.element(data2write$subid, td_Rep_matched$subid)
data2write$dataset[mask] = "Replication"

print(table(data2write$dataset, data2write$subgrp))
##              
##               RRB_over_SC SC_equal_RRB SC_over_RRB  TD
##   Discovery            10           59          64 121
##   Replication           3           62          68 122
fname2save = here(sprintf("asd_subgrp_data_rsfmri_ALL_DSM5_diffzscoreGrps_z%s.csv",as.character(z_thresh)))
write.csv(data2write,file = fname2save)

Plot mean FD and hypothesis test

df2use = subset(data2write, data2write$subgrp!="RRB_over_SC")

p = ggplot(data = df2use, aes(x = subgrp, y =meanFD, colour = subgrp)) + facet_wrap(. ~ dataset)
p = p + geom_jitter() + geom_boxplot(fill = NA, colour = "#000000", outlier.shape = NA) + guides(colour = FALSE)
p = p + xlab("Group") + ylab("Mean FD")
p

# Hypothesis test
df2use = data2write
y_var = "meanFD"
# construct linear model
# mixed-effect model: site as random factor, all other covariates as fixed factors
fx_form = as.formula(sprintf("%s ~ %s",y_var,"subgrp"))
rx_form = as.formula(sprintf("~ 1|%s","Centre"))
mod2use = eval(substitute(lme(fixed = fx_form, random = rx_form, data = df2use, na.action = na.omit)))

# run ANOVA
res = anova(mod2use)
res
##             numDF denDF   F-value p-value
## (Intercept)     1   501 116.97937  <.0001
## subgrp          3   501   1.18906  0.3133
# Discovery
df4mod = subset(df2use,df2use$dataset=="Discovery" & df2use$subgrp!="RRBoverSC")
# construct linear model
# mixed-effect model: site as random factor, all other covariates as fixed factors
fx_form = as.formula(sprintf("%s ~ %s",y_var,"subgrp"))
rx_form = as.formula(sprintf("~ 1|%s","Centre"))
mod2use = eval(substitute(lme(fixed = fx_form, random = rx_form, data = df4mod, na.action = na.omit)))

# run ANOVA
res = anova(mod2use)
res
##             numDF denDF   F-value p-value
## (Intercept)     1   246 143.96436  <.0001
## subgrp          3   246   2.28095  0.0798
# Replication
df4mod = subset(df2use,df2use$dataset=="Replication" & df2use$subgrp!="RRBoverSC")
# construct linear model
# mixed-effect model: site as random factor, all other covariates as fixed factors
fx_form = as.formula(sprintf("%s ~ %s",y_var,"subgrp"))
rx_form = as.formula(sprintf("~ 1|%s","Centre"))
mod2use = eval(substitute(lme(fixed = fx_form, random = rx_form, data = df4mod, na.action = na.omit)))

# run ANOVA
res = anova(mod2use)
res
##             numDF denDF  F-value p-value
## (Intercept)     1   247 68.92186  <.0001
## subgrp          3   247  0.09283  0.9639

Plot ADOS scores in the subtypes and also run a hypothesis test

df2use = subset(data2write, data2write$subgrp!="TD")
df2use$ados_2_SA_CSS[df2use$ados_2_SA_CSS==999] = NA
df2use$ados_2_RRB_CSS[df2use$ados_2_RRB_CSS==999] = NA

df4plot = melt(df2use,
               id.vars = c("subid","subgrp"), 
               measure.vars = c("ados_2_SA_CSS","ados_2_RRB_CSS"))
df4plot$variable = as.character(df4plot$variable)
df4plot$variable[df4plot$variable=="ados_2_SA_CSS"] = "SA"
df4plot$variable[df4plot$variable=="ados_2_RRB_CSS"] = "RRB"
df4plot$variable = factor(df4plot$variable)

p = ggplot(data = df4plot, aes(x = variable, 
                               y = value, 
                               colour = subgrp, 
                               group = subid)) + facet_grid(. ~ subgrp)
p = p + geom_point(shape=1) + geom_line(alpha = 0.2) + guides(color=FALSE)
p = p + ylab("ADOS CSS") + xlab("ADOS Subscale")
ggsave(filename = file.path(plotpath, sprintf("final_EUAIMS_ADOS_jitterplot_z%s.pdf",as.character(z_thresh))))
p

# hypothsis test on ADOS SA CSS
y_var = "ados_2_SA_CSS"
df2use = subset(data2write, !is.element(data2write$subgrp,c("TD")))

# Hypothesis test
# construct linear model
# mixed-effect model: site as random factor, all other covariates as fixed factors
fx_form = as.formula(sprintf("%s ~ %s",y_var,"subgrp"))
rx_form = as.formula(sprintf("~ 1|%s","Centre"))
mod2use = eval(substitute(lme(fixed = fx_form, random = rx_form, data = df2use, na.action = na.omit)))

# run ANOVA
res = anova(mod2use)
res
##             numDF denDF  F-value p-value
## (Intercept)     1   260 9.435649  0.0024
## subgrp          2   260 0.369413  0.6915
# hypothsis test on ADOS RRB CSS
y_var = "ados_2_RRB_CSS"
df2use = subset(data2write, !is.element(data2write$subgrp,c("TD")))

# Hypothesis test
# construct linear model
# mixed-effect model: site as random factor, all other covariates as fixed factors
fx_form = as.formula(sprintf("%s ~ %s",y_var,"subgrp"))
rx_form = as.formula(sprintf("~ 1|%s","Centre"))
mod2use = eval(substitute(lme(fixed = fx_form, random = rx_form, data = df2use, na.action = na.omit)))

# run ANOVA
res = anova(mod2use)
res
##             numDF denDF  F-value p-value
## (Intercept)     1   260 9.613633  0.0021
## subgrp          2   260 0.361243  0.6972

Descriptive stats

df2use = data2write[,c("subid","Diagnosis","dataset","Centre","meanFD","age","sex","subgrp","viq_all","piq_all","fsiq4_all","ados_2_SA_CSS","ados_2_RRB_CSS","SRS_tscore","SRS_tscore_self")]
mask = df2use==999
df2use[mask] = NA
df2use$age = df2use$age/365

cols2use = c("dataset","subgrp","age","meanFD","viq_all","piq_all","fsiq4_all","ados_2_SA_CSS","ados_2_RRB_CSS","SRS_tscore","SRS_tscore_self")
res=describeBy(x = df2use[,cols2use], group = c("subgrp"))
res
## 
##  Descriptive statistics by group 
## group: RRB_over_SC
##                 vars  n   mean    sd median trimmed   mad   min    max range
## dataset*           1 13    NaN    NA     NA     NaN    NA   Inf   -Inf  -Inf
## subgrp*            2 13    NaN    NA     NA     NaN    NA   Inf   -Inf  -Inf
## age                3 13  17.21  6.17  19.56   17.44  6.41  7.89  24.03 16.14
## meanFD             4 13   0.26  0.28   0.18    0.19  0.10  0.06   1.14  1.09
## viq_all            5 13 101.37 19.31  99.00  100.17 16.31 73.00 143.00 70.00
## piq_all            6 13 103.15 19.92 100.00  102.64 16.31 64.00 148.00 84.00
## fsiq4_all          7 13 102.46 18.80 102.00  100.91 13.34 74.00 148.00 74.00
## ados_2_SA_CSS      8 13   4.54  2.79   4.00    4.45  4.45  1.00   9.00  8.00
## ados_2_RRB_CSS     9 13   4.00  3.51   1.00    3.82  0.00  1.00   9.00  8.00
## SRS_tscore        10  9  71.22  9.44  72.00   71.22  4.45 58.00  90.00 32.00
## SRS_tscore_self   11  5  60.60  7.13  63.00   60.60  5.93 49.00  67.00 18.00
##                  skew kurtosis   se
## dataset*           NA       NA   NA
## subgrp*            NA       NA   NA
## age             -0.25    -1.80 1.71
## meanFD           2.37     4.77 0.08
## viq_all          0.43    -0.56 5.36
## piq_all          0.30     0.25 5.52
## fsiq4_all        0.73     0.30 5.21
## ados_2_SA_CSS    0.16    -1.65 0.77
## ados_2_RRB_CSS   0.32    -1.87 0.97
## SRS_tscore       0.38    -0.60 3.15
## SRS_tscore_self -0.66    -1.43 3.19
## ------------------------------------------------------------ 
## group: SC_equal_RRB
##                 vars   n   mean    sd median trimmed   mad   min    max range
## dataset*           1 121    NaN    NA     NA     NaN    NA   Inf   -Inf  -Inf
## subgrp*            2 121    NaN    NA     NA     NaN    NA   Inf   -Inf  -Inf
## age                3 121  16.43  5.87  15.73   16.09  5.89  7.08  30.28 23.20
## meanFD             4 121   0.26  0.39   0.18    0.20  0.12  0.04   3.95  3.91
## viq_all            5 119  99.98 17.03 102.00  100.43 16.31 61.00 136.00 75.00
## piq_all            6 119 101.73 19.08 102.80  102.55 19.57 52.00 142.00 90.00
## fsiq4_all          7 120 101.13 17.29 104.50  101.92 18.53 60.00 131.00 71.00
## ados_2_SA_CSS      8 118   5.92  2.54   6.00    5.98  2.97  1.00  10.00  9.00
## ados_2_RRB_CSS     9 118   5.04  2.62   5.50    5.04  1.48  1.00  10.00  9.00
## SRS_tscore        10 107  68.68 11.80  69.00   68.67 14.83 43.00  90.00 47.00
## SRS_tscore_self   11  61  61.11  9.50  61.00   60.98 10.38 43.00  89.00 46.00
##                  skew kurtosis   se
## dataset*           NA       NA   NA
## subgrp*            NA       NA   NA
## age              0.46    -0.58 0.53
## meanFD           7.34    64.00 0.04
## viq_all         -0.25    -0.65 1.56
## piq_all         -0.40    -0.35 1.75
## fsiq4_all       -0.38    -0.56 1.58
## ados_2_SA_CSS   -0.25    -0.95 0.23
## ados_2_RRB_CSS  -0.36    -0.81 0.24
## SRS_tscore      -0.01    -0.96 1.14
## SRS_tscore_self  0.23    -0.19 1.22
## ------------------------------------------------------------ 
## group: SC_over_RRB
##                 vars   n  mean    sd median trimmed   mad   min    max  range
## dataset*           1 132   NaN    NA     NA     NaN    NA   Inf   -Inf   -Inf
## subgrp*            2 132   NaN    NA     NA     NaN    NA   Inf   -Inf   -Inf
## age                3 132 16.25  5.07  15.88   16.08  5.68  7.48  29.40  21.92
## meanFD             4 132  0.25  0.25   0.16    0.19  0.11  0.03   1.59   1.55
## viq_all            5 128 96.85 19.56  99.00   97.22 20.00 50.91 142.00  91.09
## piq_all            6 130 99.07 22.02 103.00  100.27 21.46 44.03 150.00 105.97
## fsiq4_all          7 130 98.33 19.57 102.29   99.04 19.63 59.00 143.00  84.00
## ados_2_SA_CSS      8 127  6.24  2.72   6.00    6.35  2.97  1.00  10.00   9.00
## ados_2_RRB_CSS     9 127  4.59  2.77   5.00    4.49  2.97  1.00  10.00   9.00
## SRS_tscore        10 116 73.93 11.87  76.00   74.72 13.34 44.00  95.00  51.00
## SRS_tscore_self   11  57 63.63 11.11  62.00   62.91  8.90 40.00  94.00  54.00
##                  skew kurtosis   se
## dataset*           NA       NA   NA
## subgrp*            NA       NA   NA
## age              0.33    -0.55 0.44
## meanFD           2.80     8.91 0.02
## viq_all         -0.18    -0.48 1.73
## piq_all         -0.41    -0.48 1.93
## fsiq4_all       -0.30    -0.81 1.72
## ados_2_SA_CSS   -0.31    -1.05 0.24
## ados_2_RRB_CSS  -0.15    -1.23 0.25
## SRS_tscore      -0.46    -0.50 1.10
## SRS_tscore_self  0.62     0.41 1.47
## ------------------------------------------------------------ 
## group: TD
##                 vars   n   mean    sd median trimmed   mad   min    max  range
## dataset*           1 243    NaN    NA     NA     NaN    NA   Inf   -Inf   -Inf
## subgrp*            2 243    NaN    NA     NA     NaN    NA   Inf   -Inf   -Inf
## age                3 243  16.84  5.66  16.58   16.63  6.57  6.89  29.84  22.95
## meanFD             4 243   0.20  0.34   0.13    0.15  0.07  0.03   4.60   4.58
## viq_all            5 241 104.27 18.62 106.00  105.33 14.83 45.00 160.00 115.00
## piq_all            6 241 105.36 18.92 107.00  107.07 16.31 49.00 147.00  98.00
## fsiq4_all          7 241 105.33 17.70 108.18  107.16 13.61 50.00 142.00  92.00
## ados_2_SA_CSS      8   0    NaN    NA     NA     NaN    NA   Inf   -Inf   -Inf
## ados_2_RRB_CSS     9   0    NaN    NA     NA     NaN    NA   Inf   -Inf   -Inf
## SRS_tscore        10 133  47.54  9.34  44.00   45.82  4.45 37.00  90.00  53.00
## SRS_tscore_self   11 132  47.50  5.90  46.00   46.72  5.19 39.00  69.00  30.00
##                  skew kurtosis   se
## dataset*           NA       NA   NA
## subgrp*            NA       NA   NA
## age              0.28    -0.74 0.36
## meanFD           9.47   111.25 0.02
## viq_all         -0.55     0.92 1.20
## piq_all         -0.80     0.53 1.22
## fsiq4_all       -0.96     1.13 1.14
## ados_2_SA_CSS      NA       NA   NA
## ados_2_RRB_CSS     NA       NA   NA
## SRS_tscore       1.85     3.60 0.81
## SRS_tscore_self  1.18     1.39 0.51

Model age differences

y_var = "age"

# df4mod = subset(df2use,df2use$subgrp!="RRBoverSC")
df4mod = df2use
# construct linear model
# mixed-effect model: site as random factor, all other covariates as fixed factors
fx_form = as.formula(sprintf("%s ~ %s",y_var,"subgrp"))
rx_form = as.formula(sprintf("~ 1|%s","Centre"))
mod2use = eval(substitute(lme(fixed = fx_form, random = rx_form, data = df4mod, na.action = na.omit)))

# run ANOVA
res = anova(mod2use)
res
##             numDF denDF  F-value p-value
## (Intercept)     1   501 796.1510  <.0001
## subgrp          3   501   0.5653  0.6381
# Discovery
df4mod = subset(df2use,df2use$dataset=="Discovery" & df2use$subgrp!="RRBoverSC")
# construct linear model
# mixed-effect model: site as random factor, all other covariates as fixed factors
fx_form = as.formula(sprintf("%s ~ %s",y_var,"subgrp"))
rx_form = as.formula(sprintf("~ 1|%s","Centre"))
mod2use = eval(substitute(lme(fixed = fx_form, random = rx_form, data = df4mod, na.action = na.omit)))

# run ANOVA
res = anova(mod2use)
res
##             numDF denDF  F-value p-value
## (Intercept)     1   246 601.9868  <.0001
## subgrp          3   246   0.4835  0.6941
# Replication
df4mod = subset(df2use,df2use$dataset=="Replication" & df2use$subgrp!="RRBoverSC")
# construct linear model
# mixed-effect model: site as random factor, all other covariates as fixed factors
fx_form = as.formula(sprintf("%s ~ %s",y_var,"subgrp"))
rx_form = as.formula(sprintf("~ 1|%s","Centre"))
mod2use = eval(substitute(lme(fixed = fx_form, random = rx_form, data = df4mod, na.action = na.omit)))

# run ANOVA
res = anova(mod2use)
res
##             numDF denDF  F-value p-value
## (Intercept)     1   247 500.7472  <.0001
## subgrp          3   247   0.5791  0.6292

Model VIQ differences

y_var = "viq_all"

# df4mod = subset(df2use,df2use$subgrp!="RRBoverSC")
df4mod = df2use
# construct linear model
# mixed-effect model: site as random factor, all other covariates as fixed factors
fx_form = as.formula(sprintf("%s ~ %s",y_var,"subgrp"))
rx_form = as.formula(sprintf("~ 1|%s","Centre"))
mod2use = eval(substitute(lme(fixed = fx_form, random = rx_form, data = df4mod, na.action = na.omit)))

# run ANOVA
res = anova(mod2use)
res
##             numDF denDF   F-value p-value
## (Intercept)     1   493 2001.2662  <.0001
## subgrp          3   493    3.4812  0.0158
# Discovery
df4mod = subset(df2use,df2use$dataset=="Discovery" & df2use$subgrp!="RRBoverSC")
# construct linear model
# mixed-effect model: site as random factor, all other covariates as fixed factors
fx_form = as.formula(sprintf("%s ~ %s",y_var,"subgrp"))
rx_form = as.formula(sprintf("~ 1|%s","Centre"))
mod2use = eval(substitute(lme(fixed = fx_form, random = rx_form, data = df4mod, na.action = na.omit)))

# run ANOVA
res = anova(mod2use)
res
##             numDF denDF   F-value p-value
## (Intercept)     1   242 2274.5550  <.0001
## subgrp          3   242    2.0925  0.1018
# Replication
df4mod = subset(df2use,df2use$dataset=="Replication" & df2use$subgrp!="RRBoverSC")
# construct linear model
# mixed-effect model: site as random factor, all other covariates as fixed factors
fx_form = as.formula(sprintf("%s ~ %s",y_var,"subgrp"))
rx_form = as.formula(sprintf("~ 1|%s","Centre"))
mod2use = eval(substitute(lme(fixed = fx_form, random = rx_form, data = df4mod, na.action = na.omit)))

# run ANOVA
res = anova(mod2use)
res
##             numDF denDF   F-value p-value
## (Intercept)     1   243 1478.4458  <.0001
## subgrp          3   243    1.9607  0.1205

Model PIQ differences

y_var = "piq_all"

# df4mod = subset(df2use,df2use$subgrp!="RRBoverSC")
df4mod = df2use
# construct linear model
# mixed-effect model: site as random factor, all other covariates as fixed factors
fx_form = as.formula(sprintf("%s ~ %s",y_var,"subgrp"))
rx_form = as.formula(sprintf("~ 1|%s","Centre"))
mod2use = eval(substitute(lme(fixed = fx_form, random = rx_form, data = df4mod, na.action = na.omit)))

# run ANOVA
res = anova(mod2use)
res
##             numDF denDF   F-value p-value
## (Intercept)     1   495 1454.3107  <.0001
## subgrp          3   495    1.4141  0.2378
# Discovery
df4mod = subset(df2use,df2use$dataset=="Discovery" & df2use$subgrp!="RRBoverSC")
# construct linear model
# mixed-effect model: site as random factor, all other covariates as fixed factors
fx_form = as.formula(sprintf("%s ~ %s",y_var,"subgrp"))
rx_form = as.formula(sprintf("~ 1|%s","Centre"))
mod2use = eval(substitute(lme(fixed = fx_form, random = rx_form, data = df4mod, na.action = na.omit)))

# run ANOVA
res = anova(mod2use)
res
##             numDF denDF   F-value p-value
## (Intercept)     1   242 1860.9510  <.0001
## subgrp          3   242    1.5502  0.2022
# Replication
df4mod = subset(df2use,df2use$dataset=="Replication" & df2use$subgrp!="RRBoverSC")
# construct linear model
# mixed-effect model: site as random factor, all other covariates as fixed factors
fx_form = as.formula(sprintf("%s ~ %s",y_var,"subgrp"))
rx_form = as.formula(sprintf("~ 1|%s","Centre"))
mod2use = eval(substitute(lme(fixed = fx_form, random = rx_form, data = df4mod, na.action = na.omit)))

# run ANOVA
res = anova(mod2use)
res
##             numDF denDF   F-value p-value
## (Intercept)     1   245 1177.2195  <.0001
## subgrp          3   245    0.9235    0.43

Model FIQ differences

y_var = "fsiq4_all"

# df4mod = subset(df2use,df2use$subgrp!="RRBoverSC")
df4mod = df2use
# construct linear model
# mixed-effect model: site as random factor, all other covariates as fixed factors
fx_form = as.formula(sprintf("%s ~ %s",y_var,"subgrp"))
rx_form = as.formula(sprintf("~ 1|%s","Centre"))
mod2use = eval(substitute(lme(fixed = fx_form, random = rx_form, data = df4mod, na.action = na.omit)))

# run ANOVA
res = anova(mod2use)
res
##             numDF denDF   F-value p-value
## (Intercept)     1   496 2055.5837  <.0001
## subgrp          3   496    2.6133  0.0506
# Discovery
df4mod = subset(df2use,df2use$dataset=="Discovery" & df2use$subgrp!="RRBoverSC")
# construct linear model
# mixed-effect model: site as random factor, all other covariates as fixed factors
fx_form = as.formula(sprintf("%s ~ %s",y_var,"subgrp"))
rx_form = as.formula(sprintf("~ 1|%s","Centre"))
mod2use = eval(substitute(lme(fixed = fx_form, random = rx_form, data = df4mod, na.action = na.omit)))

# run ANOVA
res = anova(mod2use)
res
##             numDF denDF   F-value p-value
## (Intercept)     1   244 2810.9361  <.0001
## subgrp          3   244    2.2322  0.0851
# Replication
df4mod = subset(df2use,df2use$dataset=="Replication" & df2use$subgrp!="RRBoverSC")
# construct linear model
# mixed-effect model: site as random factor, all other covariates as fixed factors
fx_form = as.formula(sprintf("%s ~ %s",y_var,"subgrp"))
rx_form = as.formula(sprintf("~ 1|%s","Centre"))
mod2use = eval(substitute(lme(fixed = fx_form, random = rx_form, data = df4mod, na.action = na.omit)))

# run ANOVA
res = anova(mod2use)
res
##             numDF denDF   F-value p-value
## (Intercept)     1   244 1455.9095  <.0001
## subgrp          3   244    1.3255  0.2666